information delivery under endogenous communication anushka chawla, devika lakhote, alokik mishra,...

INFORMATION DELIVERY UNDER ENDOGENOUSCOMMUNICATION: EXPERIMENTAL EVIDENCE FROM THE

INDIAN DEMONETIZATION

ABHIJIT BANERJEEı, EMILY BREZA§, ARUN G. CHANDRASEKHAR‡,AND BENJAMIN GOLUB†

Abstract. To inform the public, how should policymakers disseminate informa-tion: by broadcasting widely (e.g., via mass media), or letting word spread from asmall number of initially informed “seed” individuals? While conventional wisdomsuggests more information is better, we argue that it may not be when participa-tion in social learning is endogenous. In a field experiment during the chaotic 2016Indian demonetization, we vary how information is delivered to villages: the num-ber informed (broadcasting versus seeding) and whether it is common knowledgewho is informed. Our results are consistent with four predictions of a signalingmodel in which, endogenously, people are more willing to engage in learning aboutinformation considered scarcer. First, without common knowledge, broadcastingdoes cause more conversations than seeding. Second, under seeding, adding com-mon knowledge increases conversations. Third, holding common knowledge fixed,broadcasting decreases conversations relative to seeding. Fourth, when broadcast-ing, adding common knowledge also decreases conversations. Moreover, the treat-ments that increase (decrease) conversations also lead to better (worse) knowledgeand incentivized decisions.

Date: This Version: October 2, 2017.We thank Mohammad Akbarpour, Pascaline Dupas, Matt Jackson, and Ben Olken for helpful com-ments and conversations. We thank Pooja Suri, Arnesh Chowdhury, Piyush Tank, Tithee Mukhopad-hyay, Anushka Chawla, Devika Lakhote, Alokik Mishra, and Shambhavi Priyam for exceptionalresearch assistance. Stanford IRB-35996.ıDepartment of Economics, MIT; NBER; JPAL.§Department of Economics, Harvard University; NBER; JPAL.‡Department of Economics, Stanford University; NBER; JPAL.§Department of Economics, Harvard University.

1

INFORMATION DELIVERY UNDER ENDOGENOUS COMMUNICATION 2

1. Introduction

How should a policymaker deliver information to a community? In practice, thereare two commonly used, very di�erent, strategies: (1) broadcasting information widelyto all (e.g., radio, television, newspaper, or a Twitter feed) and (2) delivering infor-mation to a select few “seed” individuals and relying on subsequent di�usion (e.g.,viral marketing, extension services, or the introduction of microcredit)1.Holding fixedagents’ participation in learning, informing more people should always be better.More signals would be received and aggregated throughout society. However, in fact,participation in learning, especially social learning, may be a�ected by the deliverystrategy used. In order to know whether broadcasting or seeding is the best strategyto follow, we first need to understand this e�ect.

To consider how delivery strategies may a�ect participation, it is helpful to ex-amine how these strategies di�er on dimensions other than the number of peopleinformed. One key dimension is individuals’ beliefs, not just about the facts beingcommunicated, but about others’ access to the information. With broadcasts, indi-viduals consider it likely that most others have access to certain information, andthat everyone understands this. Thus, in particular, a typical individual knows thatothers believe that he has access to information that has been broadcast. Meanwhile,with targeted seeding, this may or may not be true. In viral marketing campaigns,for example, there is no presumption that the community knows how many or whichindividuals were initially informed. Alternately, in settings such as model farms,agricultural extension agents may inform the community who is receiving the farm-ing information. Thus, there are two basic dimensions of delivery strategies thatguide our work: (1) the number of people to whom information is delivered and (2)the nature and extent of knowledge about the delivery mechanism.

When there is endogenous participation in social learning, these two dimensionscan interact in interesting ways. Meta-knowledge about the information deliverymechanism has one obvious consequence—namely that, when it is known who isinformed, an agent who needs clarification knows that there is information out thereand even where to look for it. But there are also less obvious e�ects. Imagine asituation in which capable individuals can learn everything they need to know fromthe initial delivery of information, as long as they receive it, but less capable ones

1See, e.g., Leskovec et al. (2007); Ryan and Gross (1943); Conley and Udry (2010); Banerjee et al.(2013); Beaman et al. (2016); Cai et al. (2015).


need help to process the information or understand its relevance. In prior work,Chandrasekhar, Golub, and Yang (2017) consider such a situation theoretically andin an experiment, and show that signaling concerns arise endogenously and detercommunication: those who especially need to participate in social learning choosenot to, in order to avoid revealing an attribute correlated with that need. In adi�erent context, the theory and experiments of Bursztyn, Egorov, and Jensen (2016)show that related signaling concerns a�ect willingness to seek help in an educationalcontext.2

Such forces have immediate consequences for dissemination strategies. If an agentknows that others believe he has received information, then in the situation of theprevious paragraph, he may be less willing to raise questions with others. But ifthe information is thought to be scarce, there is no negative signal in not havinginternalized it already. The role of ability in the above story may also be replacedor augmented by other considerations, such as a notion that it is more polite, ormore interesting, to discuss issues of common interest about which information isscarce, than to go over statements that have been made (nearly) common knowledge.All these forces will make engagement responsive—indeed, negatively a�ected—by aperception of a broad announcement.

This line of thinking about individual behavior leads us to four predictions aboutsocial learning at the community level: (1) If few signals are provided to a community,then adding common knowledge of who has the information should increase engage-ment in social learning, because it helps people find the information without exposingthem to any negative signaling concerns; (2) If there is no common knowledge of thedelivery mechanism, going from seeding the information narrowly to broadcastingshould increase engagement in social learning; (3) if there is common knowledge ofthe delivery mechanism, then going from seeding the information narrowly to broad-casting should actually reduce engagement in social learning; (4) if information isbroadcast to all, not having common knowledge of this fact can lead to more partic-ipation in social learning than when there is common knowledge.

Motivated by such considerations, we conducted a field experiment in India approx-imately six weeks after Prime Minister Narendra Modi announced the demonetization

2Concerns of this sort come out not only in theory and experiments, but also in survey data: Chan-drasekhar et al. (2017) report that 88% of households in their sample felt there was a limit on howmuch they could ask members of their community for information about farming, health, or finances.In 64% of these cases, the respondents said they refrained from engaging in learning because theydid not want to appear weak or uninformed.


of all Rs. 500 and Rs. 1000 notes. The policy was unexpected and far-reaching, af-fecting 86% of India’s currency. While there was near-universal awareness of thepolicy in India, its chaotic implementation, with over 50 rule changes in a seven weekperiod led to widespread confusion and misinformation (see Appendix A). For exam-ple, in our sample, 15% thought that the Rs. 10 coin was being demonetized and25% did not understand that over-the-counter exchange was no longer available asan option. Thus, the period following the policy announcement constitutes a fittinglaboratory with real stakes for studying information dissemination strategies availableto policy-makers.

In over 200 villages, in the ten days (starting on December 21, 2016) prior to thebank deposit deadline, we randomly varied how we provided information to villages.The experiment varied at the village level (1) whether information was provided to allhouseholds or just five seed households; (2) whether there was common knowledge ofthe delivery mechanism; (3) whether 24 facts or 2 facts were provided. The informa-tion we provided always consisted of a list of facts in a short pamphlet and the samepamphlet was provided to all households who received information in that village.The facts came directly from the Reserve Bank of India’s circulars, and thus repre-sent the information that the policy-makers themselves chose to communicate to thepublic. We then returned to the villages to collect our outcome data, approximatelythree days after the intervention.

Our outcome data measured engagement in social learning, policy knowledge, andchoice in an incentivized decision. For engagement in social learning, we asked howmany conversations villagers had about demonetization over the prior three days. Forknowledge, we asked about a number of facts. For incentivized choice, we asked thesubjects to select one of the three following options: (a) a Rs. 500 note (worth 2.5days’ wage) redeemable at the end of the day, (b) an IOU for Rs. 200 in Rs. 200notes una�ected by demonetization redeemable 3-5 days later, and (c) an IOU fordal (pigeon peas) worth Rs. 200, again redeemable 3-5 days later. At the time of thechoice, subjects still had time to deposit the Rs. 500 note at the bank, no questionsasked, for the cost of going to the bank.

A naive, frictionless perspective would say that providing information to all house-holds in a village (a 10-fold increase), providing more meta-information (commonknowledge), and providing more information per household (a 12-fold increase) shouldlead to more conversations, more knowledge being aggregated, and better decision-making on the part of our subjects. In sharp relief, a perspective that takes the


endogenous participation into account along the lines outlined above, has rather dif-ferent predictions. Namely, that adding common knowledge to a strategy where onlyfew are seeded should lead to more participation in social learning and presumablybetter aggregation. Meanwhile, if there is to be common knowledge, adding moresignals going from 5 households being informed to 100% being informed should actu-ally reduce the number of conversations. Whether this amounts to reduced learningas well is an empirical question. Finally, if information is to be broadcast, havingno common knowledge can lead to more engagement in social learning than havingcommon knowledge, since in the latter case it is known that one is supposed to havereceived signals as well.

Our core results are as follows. First we look at endogenous participation in sociallearning. Adding common knowledge to a seeding strategy adds considerable value:going from (Seed, No CK) to (Seed, CK) increases the number of conversations by103% (p = 0.04). Furthermore, if we look at strategies that use common knowledge,going from seeding 5 households to broadcasting to 100% of households, a 10-foldincrease in the number of households informed, leads to a 61% decline in the volumeof conversations (p = 0.029). However, in contrast, going from seeding to broadcastwithout common knowledge leads to an increase in the number of conversations by113% (p = 0.048). Moreover, turning to broadcast strategies, if we begin with broad-cast without common knowledge (which is not a feasible policy solution) and comparethat to broadcast with common knowledge, there is a 63% decline in the volume ofconversations (p = 0.02). That adding common knowledge has opposite e�ects acrossseeding and broadcast strategies, and that providing information to more people un-der common knowledge leads to less engagement in social learning, is consistent withthe sort of endogenous participation in learning model sketched above.

Second, we turn to whether the changes in endogenous participation in learningcorrespond to changes in knowledge. This is primarily an empirical question. To seethis, observe that even though there is less conversation happening in (Broadcast,CK) as compared to (Seed, CK), 10-times the number of households received infor-mation in the broadcast treatments, so it is entirely possible that they still learnedmore. On the other hand, if this is the kind of setting in which aggregation needsto happen to fight rumors and clean up noisy perceptions, then the social learningcomponent may be very valuable. We find that the results are consistent with so-cial learning being an important component of belief formation. Going from (Seed,No CK) to (Seed, CK) decreases the error rate on our knowledge survey by 7.3%


(p = 0.0142). Furthermore, conditional on strategies that use common knowledge,going from seeding to broadcast leads to a 5.6% increase in the error rate on ourknowledge survey (p = 0.062). This shows that even though all households receivesignals instead of merely five, the amount of knowledge for a random household is less,not more, suggesting that aggregation matters. The exact opposite happens whengoing from seeding to broadcasting, when there is no common knowledge. Turning tobroadcast strategies, adding common knowledge leads to a 4.6% increase in the errorrate, though the e�ect is not statistically significant (p = 0.17).

Third, when we look at whether subjects choose the Rs. 500 note, which at thattime was still accepted for deposit by banks, or an IOU worth Rs. 200 in cash or inkind to be paid in 3-5 days, we again see a similar pattern. Going from (Seed, No CK)to (Seed, CK) leads to a 81% increase in the probability of choosing the Rs. 500 note(p = 0.037). Going from seeding to broadcast, conditional on common knowledge,leads to a 38.5% decline in the probability of choosing the Rs. 500 note (p = 0.104).Again, in contrast, there is a 114% increase in the probability of choosing the notewhen going from seeding to broadcast, without common knowledge (p = 0.014).Looking at broadcast strategies adding common knowledge leads to a 48% decline inthe probability of choosing the Rs. 500 note (p = 0.041).

These results demonstrate the importance of considering endogenous participationin social learning when designing information dissemination strategies. They areconsistent with the idea that by varying whether others know who has or does nothave information, the incentive to participate in learning itself can be a�ected.

The remainder of the paper is organized as follows. Section 2 describes the settingand experimental design. In Section 3, we investigate the implications of the simplestfrictionless benchmark. Section 4 presents our theoretical framework where agentsendogenously choose to participate in social learning, study the Bayes-Nash equilibriaof our model, and show how social learning varies as we change whether informationis broadcast or seeded and common knowledge is or is not provided. We present ourempirical results in Section 5. Section 6 provides a discussion.

2. Experiment

2.1. Demonetization. On November 8, 2016, Indian Prime Minister Narendra Modiannounced a large-scale demonetization. At midnight after the announcement, alloutstanding Rs. 500 and Rs. 1000 notes (the “specified bank notes” or SBNs) ceased


to be legal tender. Demonetization a�ected 86% of circulating currency, and individ-uals holding SBNs had until December 30, 2016 to deposit them in a bank or posto�ce account. Modi intended for the surprise policy to curb “black money” and morebroadly to accelerate the digitization of the Indian economy. The policy a�ected al-most every household in the country, either because they held the SBNs, or throughthe cash shortages that resulted from problems in printing and distributing enoughnew bills fast enough.

The implementation of the policy was chaotic. The initial rollout revealed a num-ber of ambiguities, loopholes, and unintended outcomes. As a result, the governmentchanged the rules concerning demonetization over 50 times in the seven weeks fol-lowing the announcement. The rule changes concerned issues such as the time framefor over-the-counter exchange of SBNs, the cash withdrawal limit, the SBN depositlimit, and various exemptions—e.g., for weddings, which tend to be paid for in cash.See Appendix A for a timeline of these rule changes.

2.2. Setting. Our study takes place in 225 villages across 9 sub-districts in the stateof Odisha, India. The baseline was conducted starting December 21, 2016, the in-tervention on December 24, 2016, and the first endline ran from December 27 to30, 2016. It is important to note that the last day to legally deposit SBNs at bankbranches was December 30, 2016.

All of our villages have two or more hamlets segregated by major caste group.Typically one hamlet consists of scheduled caste and/or scheduled tribe individuals(SCST), commonly referred to as lower caste, and the other hamlet consists of generalor otherwise-backwards caste (GMOBC) individuals, commonly referred to as uppercaste. The two hamlets are typically 1/2 to 1 km apart. While the primary occupationdi�ers by caste, the majority of the people across the villages in our sample areinvolved in agriculture and agriculture-related activities. Given the hamlet structureof the study area, all of our treatments were focused in only one of the hamlets in eachvillage. Basic sample statistics are provided in Tables 1 and 2. 89% of our samplehad some kind of formal bank account, 80% were literate, and major occupationsincluded being a casual laborer (21%), domestic worker (16%), landed farmer (16%)and share-cropper (9%).

2.3. Baseline Knowledge of Demonetization Rules. Using responses from ourbaseline survey, we first explore the beliefs of villagers about the rules prior to ourintervention. While villagers almost universally understood that the Rs. 500 and Rs.


1000 notes were being taken out of circulation, we document in Table 3 that manyhouseholds had inaccurate beliefs regarding numerous other aspects of the policy. Forexample, more than 14% of the population thought (inaccurately) that the Rs. 10coin was also being taken out of circulation with the policy.3 25% of villagers believed(falsely) that, at the time of our baseline survey, they could still exchange notes atthe bank without first depositing them into an account. Moreover, only a handful ofrespondents could accurately tell us the deadline for making exchanges and only 50%of respondents could tell us that post o�ces/RBI o�ces/panchayat o�ces could beused for depositing the demonetized notes.

Our subjects were particularly uninformed about some of the economically impor-tant details, such as banks’ weekly withdrawal limits. 33% of respondents reportedthat they did not know the limits, and of those who claimed to know, only about40% of those were within Rs. 5,000 of the correct answer (Rs. 24,000). The ac-curacy of subjects’ knowledge for rules pertaining to ATM withdrawals look com-parably low, while respondents had even less knowledge about the rules governinglow-documentation accounts used by the poor. It is also important to note that thelow levels of knowledge are not due to limits to financial inclusion in the study setting.As noted before, in our sample, 89% of respondents had bank accounts (Table 2).

One might ask if it is important for poor rural farmers with limited formal savingsto understand minute details of the policy. However, one important implementationproblem associated with demonetization was that there simply were not enough notesto meet demand, which ended up a�ecting the lives of most people. For example,employers were not able to pay cash wages on time, microfinance borrowers were notable to service their loans, and demand for cash purchases at small shops fell. Even forindividuals without bank accounts, understanding the rules would have been usefulfor deciding whether to accept an IOU from an employer or customer, for example.Measuring these types of e�ects is beyond the scope of our study.

2.4. Experimental Design.

Sample. We enumerated an initial list of 276 villages which were assigned to treat-ments. We conduct our experiment in one hamlet in each village in our sample, sohalf of the villages were randomly assigned to have their GMOBC hamlet in ourexperiment and the other half to have their SCST hamlet in our experiment. We

3This specific rumor spread across much of the country and was reported in the Indian press.


randomized villages to treatments before we verified that each village met our crite-ria. Any hamlet that had fewer than 20 households was dropped from the sample,yielding a set of 221 villages. Sixteen villages were then added in a new subdistrictto increase the sample to 237.4 A baseline survey was administered only in the cho-sen hamlets, described above, from these villages. Given the rush of implementing200+ interventions in a matter of days, some field errors were made which led tonot collecting endline data (in 6 villages) or not conducting interventions nor havingendline data (in 5 of the villages). Village elders refused entry in two villages as well.Ultimately, we have a sample of 225 villages that were treated and received endlinesurveys.5

Before we describe the treatments, it is important to note that the baseline surveyalso contained a module based on Banerjee et al. (2016) (“the gossip survey”) toidentify the individuals in each treatment hamlet that were assessed by others to begood at spreading information.6,7

Treatments. All of our experimental treatment arms involved distributing pamphletswith information about demonetization to the study villages. Our goal was to spreadthe o�cial policy rules, and thus all information comes from the December 2016RBI circulars. We took this o�cial information, published by the central bank, andsubdivided it into 30 distinct policy rules. Through informal conversations in pilotvillages, we identified the 10 most useful rules for a typical villager in the study area.8

Our experimental protocol involved giving a randomly-selected set of facts to eachvillage. All individuals receiving lists of facts in a village received the same list.

4Online Appendix K repeats our main analysis dropping these new villages and shows that ourconclusions remain the same.5Unfortunately, given the rush of implementing 200+ interventions in a handful of days, in 16 ofthe villages our field team administered the intervention and endline to the wrong hamlet. Whilethis should be idiosyncratic and orthogonal to treatment, we therefore collected outcome data inthe right hamlet and redo our estimation using treatment assignment as instruments for treatmentin Online Appendix J. All our results look nearly identical.6We asked each individual “If we want to spread information about the money change policy putin place by the government recently, whom do you suggest we talk to? This person should be quickto understand and follow, spread the information widely, and explain it well to other people inthe village. Whom do you think is the best for this from your hamlet?” and we allowed them tonominate anywhere from 0 to 4 individuals.713 villages were dropped before information was even delivered because they were inaccessible tothe survey sta�. We show in Online Appendix M that this was not di�erential by treatment.8For example, one rule explained how foreigners could exchange their SBNs. This was not one ofthe “useful” facts on our list.


Our core design is a 2x2 that varies how many people got information as well aswhether there was common knowledge. We also crossed this with whether individualsreceived long or short lists of facts.9 Thus, the treatments are:

(1) Information dissemination:• Broadcast: information was provided to all households in the hamlet.• Seed: information was provided to 5 seed households in the hamlet, chosen

via the gossip survey.(2) Common knowledge:

• No common knowledge: we did not tell any subject that we were providinginformation to anyone else in the community.

• Common knowledge: we provided common knowledge of the informa-tion dissemination protocol. In “Broadcast” treatments in arm (1), everypamphlet contained a note that all other households received the samepamphlet. (Thus, if subjects understood and believed us, they had com-mon knowledge of the pamphlet’s distribution.) In the “Seed” treatments,every household received a notification that five individuals in their com-munity (who were identified) were provided information about demone-tization by us, and that the seeds were informed that we would informeveryone. Again, common knowledge of this information was inducedamong all the non-seeds. Figure 1 summarizes the design.

(3) Information volume:10

• Long: 24 facts were provided.• Short: 2 facts were provided.

Appendix B provides the total list of facts that we selected from for each pamphlet,in the manner described above and Appendix C provides examples of the pamphletswe handed out.11

2.5. Outcomes. We have three main outcomes of interest at endline: engagement insocial learning; general knowledge about facts surrounding the demonetization; and9We also attempted to get 30 villages of data where we did not intervene whatsoever and insteadonly collect endline data. We call these “status quo” villages. Unfortunately, these villages arenot entirely comparable to our core set due to implementation failures that led to violations ofrandomization. We detail this in Online Appendix L.10The Short lists of facts contained one “useful” fact, drawn uniformly at random, and a secondfact drawn uniformly at random from the remaining 29, while the Long list of facts were drawnuniformly.11Appendix G contains a version of our main analysis, looking separately at the endline knowledgeof useful facts and facts that were told to a village versus omitted from the village pamphlets.


whether the respondent selected the demonetized Rs. 500 note as opposed to an IOUpayable in 3-5 days for either Rs. 200 in non-demonetized notes or Rs. 200 in dal, astaple commodity.

First, we collected data on the volume of conversations about demonetization. Thisallows us to see whether engagement in social learning increased or decreased basedon the signal distribution and knowledge structure provided in the treatment arm.

Second, we assess knowledge of facts surrounding demonetization. We survey therespondent on 34 facts and create a simple metric of knowledge.

Third, we o�ered subjects a choice between: (a) a Rs. 500 note; (b) an IOU tobe filled in 3-5 days for Rs. 200 in two Rs. 100 notes; (c) an IOU to be filled in3-5 days for Rs. 200 worth of dal. With a 1/6 probability, subjects actually receivedthe item they chose. To implement the payment, we returned to each householdin the sample before exiting the village, rolled the die, and provided either the Rs.500 or the IOU notice.12 At the time of treatment, the cost of taking a note worthRs. 500 and depositing it into an account was not large. As noted above, 89% ofrespondents had bank accounts. Further, we document that waiting times at banksaveraged 15 minutes in the area, and the average village in our sample was within30 minutes of a bank by foot.13 At the time of our experiment, depositing the billrequired no documentation of the source of the cash. Thus, selecting Rs. 200 or theequivalent was giving up more than one day’s wages, even accounting for the travelto and time at the bank. We argue that this is evidence of confusion and measuresa willingness to pay to avoid holding on to the demonetized note in a period whereit was both legal and easy to convert (see Table 2). Further, we asked respondentswho did not choose the Rs. 500 to provide an open-ended justification at the endof the survey module. Figure 3 shows that most individuals who did not choose theRs. 500 note believed, mistakenly, that the deposit deadline had already passed. Thechoice between 200 rupees and the equivalent in dal was intended to capture generaltrust in paper currency and confusion about whether the 100 rupee bills had alsobecome demonetized. Taking the money o�ered more flexibility, and the dal waseasily available in village stores.14

12In practice, we surprised the respondents by paying the cost of going to the bank for them andgiving them the value in non-demonetized notes. Note that this was done as the last thing beforewe exited the village, after each and every subject had already locked in their responses.13At this time, there were still news reports of very long queues at banks and ATMs in other, moreurban parts of the country. However, during this period in our study area, the waits had becomemuch more manageable compared to the weeks following the policy announcement.14We explore this further in the Online Appendix G.


3. A Frictionless Benchmark

It is useful to start with a simple benchmark in which incentives to seek or passinformation are not an issue, information flows frictionlessly, and individuals processit in a Bayesian way. Consider a society where every individual receives m units ofinformation with probability q, i.i.d. Every individual can interpret all informationand pass it on to others, and seeks to be better-informed about the payo�-relevantstate. There may be a “network” constraint in terms of whom agents can pass infor-mation to and when. But, critically, the operation of this channel is independent ofthe nature of information (e.g., how many people got it). Individuals are Bayesian,with correctly specified beliefs, in using information. For canonical formalizationsof this, see Parikh and Krasucki (1990), Rosenberg, Solan, and Vieille (2009), andMueller-Frank (2013). In such a world, either an increase in m (more information perperson) or an increase in q (more people being informed) should improve decisions.This is because an increase in either parameter leads to a Blackwell improvement ininformation for the community, yielding better decision-making.

We begin by documenting in Table 4 that these two simple hypotheses from anystandard model of frictionless social learning in the data do not appear to hold in oursetting. First, do people converse more and learn more when each informed individualreceives more information (Short vs. Long)? Second, do people converse more andlearn more when more people are informed (Seed vs. Broadcast)? Panel A shows thatmore information per person does not lead to better outcomes. Providing a 12-foldincrease in the number of facts leads to a 26% decline in the number of conversations,no change in knowledge, and no change in the probability of picking Rs. 500. PanelB shows that broadcasting information to 100% of households instead of 10% leadsto no detectable change in either the number of conversations, knowledge, or in theprobability of picking Rs. 500.

To sum up, social learning can get worse when we provide a greater amount ofinformation to each person.15 Furthermore, when we provide information to ten timesthe number of people, we do not see the expected large increase in conversations andknowledge and no clear improvement in quality of decisions made. Note that this isdespite the fact that there are low levels of knowledge on average, even among seeds,15This is consistent with Carvalho and Silverman (2017), who argue that complexity can lead toworse decision-making and can lead to individuals taking dominated options. They study this issuein the context of portfolio choice.


which suggests that there is considerable scope for improvement in learning in thesecommunities. This suggests a world where mechanical benchmarks do not hold, andthere are endogenous communication choices and information frictions; we explorethese next.

4. Endogenous Participation in Social Learning

As discussed in the introduction, our focus is mainly on the the demand for infor-mation: the endogenous decisions to engage in conversations by agents who wouldlike to be informed. It is natural also to consider the supply side: whether the ini-tially informed invest e�ort in communicating. We discuss such considerations, whichcan analyzed in public goods/free-rider problems, in Appendix E, and consider theassociated predictions.

The model of reductions in seeking e�ort is related to recent work by Bursztyn et al.(2016) and especially Chandrasekhar et al. (2017) on how reputational incentivesa�ect information-seeking—indeed, the formal framework we use is borrowed fromthe latter paper.

4.1. The environment: information. The model focuses on a single decision-maker, D, and a single conversation partner A (for Adviser). The model focuses onone decision made by D: whether to seek information from A. D’s decision is denotedby

d œ {NS (No Seeking), S (Seeking)}.

D should be interpreted as a typical individual in the village, not one of thoseseeded in the seeding treatments. A should be interpreted as anyone who is likelyto have information conditional on its presence—someone in a central location (“thetown square”) who knows the interesting news in the village on a given day.16 Thismay be one of the individuals initially given information in the seeding treatments,but need not be. It will be clear in our formal assumptions where these interpretationsplay a role.

Let xi œ {0, 1} denote whether i œ {D, A} has information (in our application, in-formation about demonetization). There is a random variable p œ {CK:BC, CK:Seed, No-CK}that is publicly observed and is informative about the individuals’ information en-dowments. The following table records what, if anything, is common knowledgeconditional about each of the xi under the various values of p.

16It is important to emphasize that this needn’t be one of the seeded nodes.


p xD xA

CK:BC 1 1CK:Seed 0 1No-CK — —

In terms of interpretation, p captures what is publicly announced about the infor-mation dissemination procedure. In CK:BC, it is common knowledge that everyone(D and A in particular) have received information. In CK:Seed, it is common knowl-edge that A has information but not D. In No-CK, the public information is that itis “business as usual”: the absence of a public announcement.

The assumptions about the first two rows are stark. In practice, it is will not beexactly common knowledge that A is informed while a typical villager is uninformed.In fact, only an approximation of common knowledge will hold: it will be consideredlikely that certain A’s have heard the information, and it will be considered unlikelythat various D’s have, and these views about likelihood will be shared.

4.2. The environment: payo�s. D’s direct payo� to not seeking is denoted byV NS(xD, xA) = V NS(xD), which does not depend on A’s signal status, and the directpayo� to seeking is denoted by V S(xD, xA). These payo�s correspond to expectedpayo�s from using information to make some later decisions.

D has an ability type, a œ {H, L} (high or low). The prior belief of A that D’sability is high, a = H, is fi. In addition to the direct payo�, agents receive a perceptionpayo�. This enters D’s utility function additively: ⁄P(a = H | d), where ⁄ is a positivenumber. The idea is that D is better o� when A assesses D’s ability to be high—forexample, because A is then more likely to collaborate with S.17 D’s total payo� givenseeking decision d, is therefore,

V d(xD, xA) + ⁄P(a = H | d).

The distribution of the gain from seeking to D depends on D’s ability, a.

�(xD, xA) := V S (xD, xA) ≠ V NS (xD, xA) ≥ Fa(xD, xA).

The following are our assumptions on payo�s:

Assumption 1.

17Foundations are discussed in Chandrasekhar et al. (2017).


P1. FL(1, 1) strictly18 first-order stochastically dominates FH(1, 1), and both havefull support on the real line.19

P2. The following ranking of payo�s holds:

�(·, 0) < 0 < �(0, 1).

P3. The random variable �(0, 0)/�(0, 1) has positive density at 0.

The first assumption ensures that, when agents use a cuto� in equilibrium, seekingwhen V S (1, 1) ≠ V NS(1, 1) exceeds that cuto�, seeking will be a signal of low abil-ity when it is known or considered likely that, in D’s view (xD, xL) = (1, 1). Thesecond simply captures that the gain from seeking is negative when the adviser isignorant, and positive when he knows something while the seeker does not. The finalassumption is a regularity condition on the costs of seeking in the bad case (when Ais ignorant) relative to their benefits in the good case (when A is informed).

4.3. Environment: beliefs in the absence of common knowledge. To under-stand individuals’ incentives to seek, we need to consider both the benefits—whichdepend on whether information is out there to be found—and the costs. Given whatwe have said at the end of the last section, seeking costs depend on perceptions ofthe individual’s reason for seeking: seeking is a negative signal when the agent is be-lieved to have received information. Thus, for D’s seeking decision, it will ultimatelybe important what D believes about others’ perceptions of D’s likelihood of havinginformation.

Consider a situation corresponding to p = No-CK, i.e., the absence of a publicannouncement of how information is delivered. When there is information deliverywithout a public announcement, individuals’ beliefs about who is informed (and aboutothers’ beliefs about about who is informed) are shaped by their ordinary experienceof such situations. To motivate the key assumption we make in this section aboutthose beliefs, consider a particular description of an information delivery without apublic announcement.

For concreteness, one can think of a world where leaflets are distributed on somedays to some people in a community. First, nature decides whether informationdelivery occurs on this particular day: with probability µ it does; otherwise nobodyreceives information. Conditional on information delivery, people are informed i.i.d.with probability —.18I.e., for every cuto�, FL(1, 1) assigns a lower probability of being below that cuto�.19Full support is assumed for simplicity: wide enough support would su�ce.


Now, consider any i and j. Let — denote Pi(xj = 1 | xi = 1): i’s subjectiveprobability, given that i himself is informed, that j is informed. (All probabilitieshere are also conditioned on the public information of p = No-CK.) Then what wehave said about the way information is delivered implies that:

Ei[Pj(xi = 1 | xj) | xi = 1] = Pi(xj = 1 | xi = 1)ËPj(xi = 1 | xj = 1) + O(µ)

È

= —2 + O(—µ)

The left-hand side, in words, is i’s expectation of j’s subjective probability, conditionalon receiving information himself, that i is informed. In other words, it is a measure ofi’s second-order beliefs about information in the community. The calculation allowsus to bound this quantity. In particular, if µ Æ C—, so that social informationcampaigns are not very frequent, then the left-hand side can be bounded as O(—2).The second-order expectations about the receipt of information in the community,even from the perspective of someone who has received information, are of a smallerorder than the first-order expectations (from the same perspective).

Assumption B1 below captures the aspect of higher-order uncertainty, illustratedby the example, that will be important for us.

Assumption 2. Let — = Pi(xj = 1 | xi = 1).B1. Conditional on p = NO-CK and xD = 1, D’s expectation of PA(xD = 1 | xA)

is o(—).B2. Conditional on p = NO-CK, P(xA = 1 | xD = 0) is o(—).

Assumption B2 simply says that the probability one assesses of someone else beinginformed, conditional on not having seen information, is much smaller than the sameprobability conditional on being informed. This is also seen to be verified in theexample if µ is small (so that µ— is o(—)).

Though it did not play a role in our story, the assumption can accommodate a smallprobability of everyone receiving leaflets without there being common knowledge ofthat. As long as the probability of such an event is perceived to be small, the additionof this event will not upset our analysis.

4.4. Result. By the basic results about the model in Chandrasekhar et al. (2017),an equilibrium in cuto� strategies exists for for each realization of p and xD. Indeed,there is a number v(xD; p) so that an agent’s decision (independent of ability type) is

d = S if V S (xD, xA) ≠ V NS (xD, xA) > v(xD; p)


d = NS if V S (xD, xA) ≠ V NS (xD, xA) < v(xD; p)

Let ‡(xD; p) be the probability that a seeker (pooling across ability types) goesto seek advice. We focus on this statistic because it is what is observed in ourexperiments.

Proposition 1. Assume that P1-P3 and B1-B2 hold. For each large enough ⁄

(importance of signaling concerns), if — is small enough, the following inequalitiesamong values ‡(xD; p) hold:

‡(1; CK:BC) Æ ‡(0; No-CK) < ‡(1; No-CK) Æ ‡(0; CK:Seed).

That is, there is least seeking in broadcast with common knowledge, followed byseed without common knowledge, followed by broadcast without common knowledge,and most seeking in seed with common knowledge. The proof appears in Section D.

4.5. Application to Experiment. The theoretical model sketched above is ex-tremely simplified, but gives guidance on how demand for information, when a�ectedby signaling concerns, would respond to our treatments.

• We expect to see more seeking under Seeding and Common Knowledge ascompared with Seeding without Common Knowledge: turning on CommonKnowledge makes information easy to find and signaling concerns are notsubstantial: only those who got information have a chance to signal ability,and since few people did, there is not much signal in the decision to seek.

• We expect to see more seeking in (Broadcast, No CK) than in (Seed, No CK):note that the public signal p =NO-CK same in both. When D receives noinformation, xD = 0, then information is very unlikely to be out there (because— is low), so most agents will not find it worthwhile to seek even at a low cost.However, when xD = 1, information is more likely to be out there, and sincesmall costs have positive probability, some agents will choose to seek. Sincemany more agents receive the information in Broadcast, there will overall bemore seeking.

• (Seed, CK) features more seeking than (Broadcast, CK): the Broadcast treat-ment turns on signaling concerns, deters conversations. In either case, Dknows where to find information.

• (Broadcast, No CK) features more seeking than (Broadcast, CK). The argu-ment we made in the second bullet point implies considerable seeking when


xD = 1 and p =No-CK. As in the third bullet point, signaling concerns deterseeking in the CK case.

5. Results

5.1. Endogenous Participation in Social Learning. We begin by looking atwhich delivery mechanisms led to more or less endogenous participation in sociallearning, measured by the number of conversations the subject had over the priorthree days about demonetization.

Table 5 presents regressions of the number of conversations on treatment.20 Ineach regression, (Seed, No CK) is the omitted treatment arm. The coe�cients areadditive, so to compare (Broadcast, Common Knowledge) to the omitted category,it is necessary to add the coe�cients: CK, Broadcast, and Broadcast ◊ CK. In eachregression specification, we present the p-values for two key comparisons. The test(CK + BC x CK = 0) allows us to compare (Broadcast, CK) to (Broadcast, No CK).The test (BC + BC ◊ CK = 0) allows us to compare (Broadcast, CK) with (Seed,CK).

The outcome variable in column 1 is all conversations that the respondent was apart of. Going from (Seed, No CK) to (Seed, CK) increases the number of conversa-tions by 103% (0.65 more conversations, p = 0.04). This of course is consistent withthe model described above: a typical villager now knows that there is no expectationthat they had to know the information because it is common knowledge that theydid not receive signals and because seeds are known.21

Next we look at what happens when we compare strategies that employ commonknowledge. Going from (Seed, CK) to (Broadcast, CK), which corresponds to a10-fold increase in the number of households informed from 5 households to 100%of households, leads to a 61% decline in the volume of conversations (0.78 fewerconversations, p = 0.029). Again this is consistent with the model. Though thereare gains from seeking, because it is known that the agent himself has also directlyreceived signals, we see less engagement in social learning though many more peopleare informed.20For all of our main results, we focus on our core 2x2 treatment design, pooling across the Longand Short lists of facts. Appendix F provides the analysis separately for Long and Short informationand also discusses how one might interpret the length of the fact list through the lens of the model.21Of course, seeds may also have more of a motivation to spread information, though we do not findevidence consistent with this. See Appendix H.


In sharp contrast, when we look at (Broadcast, No CK) versus (Seed, No CK), weare comparing a situation where we provide signals to all versus just a few, but ineither case no agent knows whether or not any other agent has necessarily received asignal. We find that adding a 10-fold increase in the number of households informedleads to an increase in the volume of conversations by 113% (0.708 more conversations,p = 0.048). This is consistent with the idea that it is more likely that others have somesignals, so the return to seeking is higher for a random household with (Broadcast,No CK), whereas in the (Seed, No CK) case in the model a typical household doesn’teven know that there is something to converse about. The previous two results showthat informing more people is beneficial in the case without common knowledge, butdetrimental to conversation rate when there is common knowledge.

When we look at broadcasting strategies, going from (Broadcast, No CK) to(Broadcast, CK) corresponds to a 63% decline in the volume of conversations ( 0.84fewer conversations, p = 0.02). This is consistent with the model when stigma isstrong: in that world, the fact that it is not common knowledge that one receiveda signal allows the agent to ask questions about demonetization more freely. Thiscorresponds to a decline in participation in social learning.

That adding common knowledge has opposite e�ects across seeding and broadcaststrategies, and that providing information to more people under common knowledgeleads to less engagement in social learning, is consistent with the sort of endogenousparticipation in the learning model sketched above.

Having looked at conversations as a whole, we can next inquire about the types ofconversations. Loosely speaking, there are two types of conversations. First, there areconversations that are initiated with the sole purpose of talking about demonetization.These are of the form where either a respondent seeks information from others orwhere others seek information from the respondent. Second, there are conversationswhich were initiated organically in a di�erent setting – an unanticipated or incidentalconversation – but where the topic of demonetization may have come up. One canthink about this as nature pairing randomly two agents from the village, and thenthey can choose in this happenstance whether or not they ask each other aboutdemonetization. Note that incidental conversations comprise the vast majority, 78%,of reported conversations as one might imagine.

Columns 2 and 3 of Table 5 break up the number of conversations that the subjectparticipated in by whether they were incidental (column 2) or purposeful (column 3).When we look at incidental conversations we see that going from (Seed, No CK) to


(Seed, CK) corresponds to a 91% or a 0.447 increase, on a base of 0.49, in the volumeof conversations (column 2, p = 0.09). Then when we look at going from (Seed, CK)to (Broadcast, CK), we see a 63% decline in the volume of incidental conversations(column 2, p = 0.04). Again, this decline in conversation volume is overturned whenthere is no common knowledge: going from (Seed, No CK) to (Broadcast, No CK)leads to a 106% or 0.52 increase in conversation volume. Finally, when we go from(Broadcast, No CK) to (Broadcast, CK), there is a 66% decline in the number ofincidental conversations.

When we look at purposeful conversations, we see broadly a similar pattern, thoughthe e�ects are slightly more noisy. Comparing (Seed, No CK) to (Seed, CK) we see a150% or 0.204 increase, on a base of 0.137, in the volume of purposeful conversations.We cannot statistically distinguish (Seed, CK) and (Broadcast, CK) with p = 0.247,though going from (Broadcast, No CK) and (Broadcast, CK) corresponds to a 54%decline in conversations (p = 0.119). That our e�ects are starker in incidental con-versations makes sense because purposeful conversations are far sparser (comprisingonly 21.9% of conversations) and are likely to be inframarginal. One may be mostinclined to go to specific people in a doubt, but many if not most of the conversationsabout a topic may organically arise in everyday life.

To sum up, our results show that common knowledge matters considerably whenseeding with only a few, leading to far more conversation. We have also showntwo striking non-monotonicities consistent with our perspective. Adding commonknowledge to a broadcast delivery mechanism can discourage conversations. Further,if there is common knowledge, going from only 10% to 100% of the population beinginformed actually discourages conversations. As one may have expected, if there isno common knowledge, increasing the number informed increases conversations, incontrast. Next we ask whether this has implications for information aggregation andsubsequent choice behavior.

5.2. Information Aggregation and Choice. We present regressions in Table 6which show how knowledge and choice behavior depend on treatment cell.

In column 1, we turn to whether the changes in endogenous participation in learningcorrespond to changes in knowledge. This is primarily an empirical question. To seethis, observe that even though there is less conversation happening in (Broadcast, CK)as compared to (Seed, CK), 10-times the number of household received informationunder broadcast treatments, so it is entirely possible that they still learned more. Onthe other hand, if this is the kind of setting in which aggregation needs to happen to


fight rumors and clean up noisy perceptions, then the social learning component maybe very valuable.

We find that the results are consistent with social learning being an important com-ponent of belief formation. The outcome variable is the error rate on our knowledgemetric. In our metric, the (Seed, no CK) mean is 0.434. Going from (Seed, No CK)to (Seed, CK) decreases the error rate on our knowledge survey by 7.3% (p = 0.0142).Furthermore, conditional on strategies that use common knowledge, going from seed-ing to broadcast leads to a 5.6% increase in the error rate on our knowledge survey(p = 0.062). This is striking and shows that though 100% of households receiveinformation instead of 10%, the amount of aggregated information that a randomhousehold has at the end of the day is actually less, not more. Again, however, in thecase where there is no common knowledge, going from (Seed, No CK) to (Broadcast,No CK) actually increases information and declines the error rate by 6.4% (p = 0.05).Finally, turning to broadcast strategies, adding common knowledge leads to a 4.6%increase in the error rate, though the e�ect is not statistically significant (p = 0.174).All of this strongly suggests that aggregation plays a crucial role and that the declinein conversations is followed by a decline in knowledge. Exactly when we saw declinesin conversations do we see declines in knowledge.

In column 2, we turn to the impacts of our experimental treatments on incentivizedchoice. We look at whether subjects choose the Rs. 500 note on the spot which wasstill able to be deposited in their accounts (worth 2.5 days’ wage) or an IOU worthRs. 200 to be paid in 3-5 days. The probability of selecting the Rs. 500 note in theomitted category (Seed, No CK) is only 5.92%. Going from (Seed, No CK) to (Seed,CK) leads to a 4.8pp or an 81% increase in the probability of choosing the Rs. 500 note(p = 0.037). Meanwhile, going from seeding to broadcast, conditional on commonknowledge, leads to a 38.5% or 4.13pp decline in the probability of choosing the Rs.500 note (p = 0.104). But when there is no common knowledge, this corresponds to a6.77pp or 114% increase in the probability of choosing the Rs. 500 note (p = 0.014).Looking at broadcast strategies adding common knowledge leads to a 48% decline inthe probability of choosing the Rs. 500 note (p = 0.041).

Taken together, we have shown that broadcasting information is better than seedingin a world without common knowledge. However, increasing the number of informedhouseholds has opposite e�ects, depending on the presence of Common Knowledge.In a world without Common Knowledge, the conventional wisdom holds: increasingthe number informed encourages more conversations and better decision making.


However, under Common Knowledge, broadcasting information actually backfires,leading to worse outcomes across the board. These results are consistent with ourframework of strategic communication. One bottom line result is that seeding just fivehouseholds combined with Common Knowledge makes the outcomes indistinguishablefrom (Broadcast, No CK), where ten times as many people were seeded. And finally,and perhaps more striking, across the board either holding fixed Common Knowledge,and moving from Seed to Broadcast, or holding fixed Broadcast and moving from NoCommon Knowledge to Common Knowledge, actually reduces conversation volume,knowledge, and quality of choice.

6. Conclusion

Social learning happens in part through endogenous communication. We showthat, consistent with our framework and prior work by Chandrasekhar et al. (2017),the number of signals and the structure of common knowledge matters considerablyfor the extent of participation in social learning. When looking at targeted seeding,going from no common knowledge to common knowledge increases conversations.The exact opposite is true for broadcast strategies. And strikingly, conversationsactually decline when, holding common knowledge fixed, more people are providedinformation. Furthermore in our setting, this increase or decline in conversationvolume is met with a corresponding increase or decline in knowledge about the rulesas well as quality of choice. Thus, the success of an information intervention dependscrucially on the details of the design and how it a�ects endogenous communication.

Of the full set of experimental interventions, two consistently perform well on theconversations, knowledge, and choice dimensions and have comparable benefits to oneanother: seed with common knowledge and broadcast without common knowledge.Note, however, that broadcast, no common knowledge does not lend itself well towidespread policy scale-up. Most, if not all, broadcast technologies such as radio,television, newspaper, or political rallies contain intrinsically a common knowledgecomponent.

The results have implications for how researchers and policymakers should thinkabout the use of news media versus extension to educate individuals, and how ex-tension should be structured. The results indicate that the benefits of extensionstrategies can be magnified with common knowledge.


References

Banerjee, A., A. G. Chandrasekhar, E. Duflo, and M. O. Jackson (2013):“The di�usion of microfinance,” Science, 341, 1236498. 1

——— (2016): “Gossip: Identifying Central Individuals in a Social Network,” Work-ing Paper. 2.4

Banerjee, A., L. Iyer, and R. Somanathan (2007): “Public action for publicgoods,” Handbook of Development Economics, 4, 3117–3154. E

Banerjee, A. V. and E. S. Maskin (1996): “A Walrasian theory of money andbarter,” The Quarterly Journal of Economics, 111, 955–1005. G

Beaman, L., A. BenYishay, J. Magruder, and A. M. Mobarak (2016): “CanNetwork Theory based Targeting Increase Technology Adoption?” Working Paper.1

Bursztyn, L., G. Egorov, and R. Jensen (2016): “Cool to Be Smart or Smartto Be Cool? Understanding Peer Pressure in Education,” . 1, 4

Cai, J., A. De Janvry, and E. Sadoulet (2015): “Social networks and thedecision to insure,” American Economic Journal: Applied Economics, 7, 81–108. 1

Carvalho, L. and D. Silverman (2017): “Complexity and Sophistication,” . 15Chandrasekhar, A. G., B. Golub, and H. Yang (2017): “Signaling, Stigma,

and Silence in Social Learning,” Working Paper. 1, 2, 4, 17, 4.4, 6, EConley, T. G. and C. R. Udry (2010): “Learning about a New Technology:

Pineapple in Ghana,” The American Economic Review, 35–69. 1Galeotti, A. and S. Goyal (2010): “The Law of the Few,” American Economic

Review, 100, 1468–1492. EKiyotaki, N. and R. Wright (1989): “On money as a medium of exchange,”

Journal of Political Economy, 97, 927–954. GLeskovec, J., L. A. Adamic, and B. A. Huberman (2007): “The dynamics of

viral marketing,” ACM Transactions on the Web (TWEB), 1, 5. 1Mueller-Frank, M. (2013): “A general framework for rational learning in social

networks,” Theoretical Economics, 8, 1–40. 3Niehaus, P. (2011): “Filtered Social Learning,” Journal of Political Economy, 119,

686–720.Parikh, R. and P. Krasucki (1990): “Communication, consensus, and knowl-

edge,” Journal of Economic Theory, 52, 178–189. 3Rosenberg, D., E. Solan, and N. Vieille (2009): “Informational externalities

and emergence of consensus,” Games and Economic Behavior, 66, 979–994. 3


Ryan, B. and N. C. Gross (1943): “The di�usion of hybrid seed corn in two Iowacommunities.” Rural sociology, 8, 15. 1

Samuelson, P. A. (1958): “An exact consumption-loan model of interest with orwithout the social contrivance of money,” Journal of Political Economy, 66, 467–482. G

Wallace, N. (1980): “The Overlapping Generations Model of Fiat Money, in Modelsof Monetary Economies (J. Kareken and N. Wallace, eds.), 49(82),” Federal ReserveBank of Minneapolis. G


Figures

Quantity of Info per HH

Seed, ShortNCK

Seed, LongNCK

Broadcast, ShortNCK

Broadcast, LongNCK

Num

ber o

f HHs

Inform

ed

Quantity of Info per HH

Num

ber o

f HHs

Inform

ed

Common Knowledge: NO YES

Seed, ShortCK

Seed, LongCK

Broadcast, ShortCK

Broadcast, LongCK

Figure 1. Experimental Design

Nov. 8Demon. announced

Nov. 24Exchange facility

closed

Dec. 21-24Baseline

Dec. 23-27Information

delivery

Dec. 26-30Endline+ choice

Dec. 30Last day to

deposit SBNs

Figure 2. Intervention Timeline


0.2

.4.6

.81

fract

ion

of re

spon

dent

s

Bank is too Far Lines are too LongIllegal to hold Rs.500 Cannot deposit/use anymore

Figure 3. Why did you not choose 500?


01

2vo

lum

e of

con

vers

atio

ns

Seed, No CK Seed, CK Broadcast, No CK Broadcast, CK

(a) Volume of conversations

-.2-.1

0.1

.2St

anda

rdiz

ed e

rror r

ate


(b) Knowledge error

0.1

.2fra

c. c

hose

500


(c) Chose old 500

Figure 4. Core Experiment Outcomes


Tables

Table 1. Summary Statistics

mean sd obsFemale 0.32 (0.47) 1082SC/ST 0.50 (0.50) 1082Age 39.18 (11.88) 1079Casual laborer 0.21 (0.41) 1082Farmer: landed 0.16 (0.37) 1082Domestic work 0.16 (0.37) 1082Farmer: sharecropper 0.09 (0.29) 1082Unemployed 0.02 (0.14) 1082Bank account holder 0.89 (0.31) 1078Literate 0.80 (0.40) 1047Notes: This table gives summary statisticson the endline sample used for analysis.

Table 2. Bank Summary Statistics

median mean sd obsActual wait time at banks (mins) 10.00 11.86 (7.87) 51Perceived wait time at banks (mins) 15.00 17.06 (22.13) 32Nearest Bank (mins) 20.00 19.84 (9.88) 63Notes: This table gives actual wait time at banks near our samplevillages. We surveyed bank employees at 51 banks. It also givesperceived wait time and perceived time taken to reach the nearestbank by a sub-sample of the endline respondents.

Table 3. Baseline Error Statistics

mean sd obs10 rupees coin 0.15 (0.36) 965General currency 0.17 (0.38) 965Withdrawal limits on Jan Dhan accounts 0.87 (0.33) 965Over-the-counter exchange 0.25 (0.44) 965Weekly withdrawal limits from bank accounts 0.78 (0.41) 965Daily withdrawal limits on ATMs 0.90 (0.30) 965Exchange locations other than banks 0.50 (0.50) 966Notes: This tables gives error rates on knowledge about demoneti-zation in the baseline sample.


Table 4. Frictionless benchmark

Panel A: Short vs. Long(1) (2) (3)

OLS OLS OLSVARIABLES Volume Knowledge error Chose 500

Long -0.296 0.00692 -0.0183(0.250) (0.00946) (0.0180)[0.238] [0.465] [0.309]

Observations 1,078 1,082 1,067Short Mean 1.136 0.417 0.0954

Panel B: Seed vs. Broadcast(1) (2) (3)

OLS OLS OLSVARIABLES Volume Knowledge error Chose 500

Broadcast -0.0399 -0.00236 0.0129(0.253) (0.00936) (0.0186)[0.875] [0.802] [0.490]

Observations 1,078 1,082 1,067Seed Mean 0.998 0.418 0.0755Notes: All columns control for randomization strata(subdistrict) fixed e�ects. They also control for dateand time of entry into the village, caste category of thetreatment hamlet and distance from the village to anurban center. Respondent-level controls include age,gender, literacy and potential seed status. Standarderrors (clustered at the village-level) are reported inparentheses and p-values are reported in brackets.


Table 5. Engagement in social learning

(1) (2) (3)OLS OLS OLS

Volume of # incidental # purposefulVARIABLES conversations conversations conversations

CK 0.651 0.447 0.204(0.318) (0.262) (0.105)[0.0420] [0.0901] [0.0527]

Broadcast 0.708 0.520 0.188(0.356) (0.320) (0.127)[0.0477] [0.106] [0.142]

Broadcast ◊ CK -1.491 -1.113 -0.378(0.529) (0.442) (0.190)

[0.00535] [0.0125] [0.0482]

Observations 1,078 1,078 1,078Seed, No CK Mean 0.627 0.490 0.137CK + BC ◊ CK = 0 p-val 0.0211 0.0314 0.247BC + BC ◊ CK = 0 p-val 0.0292 0.0399 0.119Notes: All columns control for randomization strata (subdistrict) fixed ef-fects. They also control for date and time of entry into the village, castecategory of the treatment hamlet and distance from the village to an ur-ban center. Respondent-level controls include age, gender, literacy andpotential seed status. Standard errors (clustered at the village-level) arereported in parentheses and p-values are reported in brackets.


Table 6. Knowledge and decision-making

(1) (2)OLS OLS

VARIABLES Knowledge Error Chose 500

CK -0.0318 0.0480(0.0129) (0.0228)[0.0142] [0.0368]

Broadcast -0.0279 0.0677(0.0143) (0.0272)[0.0525] [0.0135]

Broadcast ◊ CK 0.0506 -0.109(0.0193) (0.0392)[0.00958] [0.00583]

Observations 1,082 1,067Seed, No CK Mean 0.434 0.0592CK + BC ◊ CK = 0 p-val 0.174 0.0409BC + BC ◊ CK = 0 p-val 0.0621 0.104Notes: All columns control for randomization strata (sub-district) fixed e�ects. They also control for date and timeof entry into the village, caste category of the treatmenthamlet and distance from the village to an urban center.Respondent-level controls include age, gender, literacy andpotential seed status. Standard errors (clustered at thevillage-level) are reported in parentheses and p-values arereported in brackets.


Appendix A. Timeline of Rule Changes

Nov-08 • Rs. 500 and Rs. 1000 notes shall have their legal tenderwithdrawn wef midnight Nov 8

• Closure of ATMs from Nov 9th to Nov 11th• All ATM free of cost of dispensation

• ATM machine withdrawal limit:Rs. 2000 per day per card (till Nov. 18th); Rs. 4000 thereafter

Nov-09 • Re-Calibration of ATMs to dispense Rs. 50 and Rs. 100 notes• Withdrawal of Rs. 2000 limit per day per card

• Cash withdrawals could be made from Banking Correspondentsand Aadhar Enabled Payment Systems

Nov-10 • Rs. 4000 or below could be exchanged for any denomination at banks• Max deposit for an account without KYC: Rs. 40000

• Cash withdrawal per day: Rs. 10,000; with a limit of Rs. 20,000in one week

Nov-13 • Limit for over the counter withdrawal: Rs. 4500• Daily withdrawal on debit cards: Rs. 2500• Weekly withdrawal limit: Rs. 24,000• Daily limit of Rs. 10,000: withdrawn• Separate queues for senior citizens and disabled

Nov-14 • Waivers of ATM customer charge

• Current account holders: Withdrawal limits Rs. 50,000with notes of mostly Rs. 2000

Nov-17 • Over the counter exchange of notes limited to Rs. 2000

• PAN card is mandatory for deposits over Rs. 50,000, oropening a bank account

Nov-20 • Withdrawal of ATM: limit unchanged at Rs. 2500

Nov-21 • Cash withdrawal for wedding: Rs. 2,50,000 for each partyfor wedding before Dec. 30th, for customers with full KYC

• 60 day extra for small borrowers to repay loan dues

• Limit of Rs. 50,000 withdrawal also extended to overdraft,cash credit account (in addition of current account - Nov-14)

• Farmers can purchase seeds with the old Rs. 500 notes

Nov-22 • Prepaid payment instruments: limit extended from Rs. 10,000to Rs. 20,000 in order to push electronic payment systems

•For wedding payments: a list must be provided with detailsof payments for anyone to whom a payment of more that 10,000is to be made for wedding purposes

Nov-23 • SBNs not allowed to deposit money in Small Saving SchemesNov-24 • No over the counter exchange of SBNs wef midnight Nov-24

•Only the old Rs. 500 notes will be accepted till Dec. 15thin the following places: government school or college fees,pre-paid mobiles, consumer co-op stores, tolls for highways

Nov-25 • Weekly withdrawal limit: Rs. 24,000 (unchanged)• Foreign citizens allowed to exchange Rs. 5000 per week till Dec 15th


Nov-28 • Relaxation in norms of withdrawal from deposit accounts of deposits made inlegal tender note wef Nov-29

Nov-29 •For account holders of Pradhan Mantri Jan Dhan Yojana:limit of Rs. 10,000 withdrawal per month for full KYCcustomers; Rs. 5000 with customers with partial KYC

Dec-02 • Aadhaar-based Authentication for Card Present Transactions

Dec-06 • Relaxation in Additional Factor of Authentication for payments upto Rs. 2000for card network provided authentication solutions

Dec-07 • Old Rs. 500 notes can only be used for purchase of railway tickets till Dec. 10thDec-08 • OTP based e-KYC allowedDec-16 • Pradhan Mantri Garib Kalyan Deposit Scheme Issued wef Dec 17

• Foreign citizens allowed to exchange Rs. 5000 per week till Dec 31st• Merchant discount rate for debit card transactions revised• No customer charges to be levied for IIMPS, UPI, USSD

Dec-19 • SBNs of more than Rs. 5000 to be accepted only once till Dec 30thto full KYC customers

Dec-21 • The limit of Rs. 5000 deposit not applicable to full KYC customersDec-26 • 60 day extra for short term crop loansDec-29 • Additional working capital for MSEsDec-30 • Closure of the scheme of exchange of Specified Bank Notes

• PPI guideline (issued Nov 22) extended• ATM machine withdrawal limit: Rs. 4500 per day per card

Dec-31 • Grace period for non-present Indians for SBN exchange at RBIJan-03 • Allocation changes to cash in rural areas

• Foreign citizens allowed to exchange Rs. 5000 per week till Jan 31Jan-16 • ATM limit extended to Rs. 10,000 per day per card

• Current account withdrawal limits extended to 1,00,000


Appendix B. List of Facts

Chapter 1:DEPOSITING ORTENDERING SPECIFIEDBANK NOTES

1. The old Rs. 500 and Rs.1000 notes will be accepted at bank branches until30/12/2016. If you deposit more than Rs. 5,000 then you will have toprovide a rationale for why you didnt deposit the notes earlier.2. You will get value for the entire volume of notes tendered at thebank branches / RBI o�ces.3. If you are not able to personally visit the branch, you may send a representativewith a written authority letter and his/her identity proof with tendering the notes.4. Banks will not be accepting the old Rs.500 and Rs. 1000 notes for deposits inSmall Saving Schemes. The deposits canbe made in Post O�ce Savings accounts.5. Quoting of PAN is mandatory in the following transactions: Deposit with a bankin cash exceeding Rs. 50,000 in a single day; Purchase of bank drafts or pay ordersor bankers cheques from a bank in cash for an amount exceeding Rs. 50,000 in asingle day; A time deposit with a Bank or a Post O�ce; Total cash depositof more than Rs. 2,50,000 during November 09 to December 30th, 2016

Chapter 2:EXCHANGINGSPECIFIED BANKNOTES

1. The over the counter exchange facility has been discontinued from the midnightof 24th November, 2016 at all banks. This means that the bank wont exchangethe notes for you anymore. You must first deposit them into an account.2. All of the old Rs.500 and Rs. 1,000 notes can be exchanged at RBI O�ces only,up to Rs.2000 per person.3. Until December 15th, 2016, foreign citizens will be allowed to exchange up toRs. 5000 per week. It is mandatory for them to have this transaction enteredin their passports.4. Separate queues will be arrangedfor Senior Citizens and Divyang persons,customers with accounts in the Bankand for customers for exchange of notes(when applicable).

Chapter 3:CASH WITHDRAWALAT BANK BRANCHES

1. The weekly limit of Rs. 20,000 for withdrawal from Bank accounts hasbeen increased to Rs. 24,000. The limit of Rs. 10,000 per day has been removed.2. RBI has issued a notification to allow withdrawals of deposits made in the validnotes (including the new notes) on or after November 29, 2016 beyond the currentlimits. The notification states that available higher denominations bank notesof Rs. 2000 and Rs. 500 are to be issued for such withdrawals as far as possible.3. Business entities having Current Accounts which are operational for last threemonths or more will be allowed to draw Rs. 50,000 per week. This can be donein a single transaction or multiple transactions.4. To protect innocent farmers and rural account holders of PMJDY from moneylaunders, temporarily banks will: (1) allow account holders with full KYC towithdraw Rs. 10,000 in a month;(2) allow account holders with limited KYC towithdraw Rs.5,000 per month, withthe maximum of Rs.10,000 from the amountdeposited through SBN after Nov 09,20165. District Central Cooperative Banks (DCCBs) will also facilitate withdrawals withthe same limits as normal banks.

Chapter 4:ATM WITHDRAWALS

1. Withdrawal limit increased to Rs. 2,500 per day for ATMs that have beenrecalibrated to fit the new bills. This will enable dispensing of lower denominationcurrency notes for about Rs.500 per withdrawal. The new Rs. 500 notescan be withdrawn2. Micro ATMs will be deployed to dispense cash against Debit/Credit cards up tothe cash limits applicable for ATMs.3. ATMs which are yet to berecalibrated, will continue to dispense Rs. 2000 tillthey are recalibrated.

Chapter 5:SPECIAL PROVISIONSFOR FARMERS

1. Farmers would be permitted to withdraw up to Rs. 25,000 per week in cashfrom their KYC compliant accounts for loans. These cash withdrawals would besubject to the normal loan limits and conditions. This facility will also applyto the Kisan Credit Cards (KCC).2. Farmers receiving payments into their bank accounts through cheque or otherelectronic means for selling their produce, will be permitted to withdraw up toRs.25,000 per week in cash. But these accounts will have to be KYC compliant.3. Farmers can purchase seeds with the old bank notes of 500 from the State orCentral Govenment Outlets, Public Sector Undertakings, National or State SeedsCorporations, Central or State Agricultural Universities and the Indian Council ofAgricultural Research (ICAR), with ID proof.

1


4. Traders registered with APMC markets/mandis will be permitted to withdrawup to Rs. 50,000 per week in cashfrom their KYC compliant accounts as in thecase of business entities.5. The last date for payment of crop insurance premium has been extended by15 days to 31st December,2016.

Chapter 6:SPECIAL PROVISIONSFOR WEDDINGS

1. In the case of a wedding, one individual from the family (parent or the person them-selves) will be able to withdraw Rs. 2,50,000 from a KYC compliant bank account.PAN details and self-declaration will have to be submitted stating only one person iswithdrawing the amount. The girls and the boys family can withdraw thisamount separately.2. The application for withdrawal for a wedding has to be accompanied by the followingdocuments: An application form; Evidence of the wedding, including the invitation card,copies of receipts for advance payments already made, such as Marriage hall booking,advance payments to caterers, etc.; A declaration from the person who has to be paid morethan Rs. 10,000 stating that they do not have a bank account, anda complete list of peoplewho have to be paid in cash and the purpose for the payment.

Chapter 7:OTHER DETAILS

1. In Odisha, Panchayat o�ces can be used for banking services in areas where banksare too far or banking facilities are not available.2. You can use NEFT/RTGS/IMPS/InternetBanking/Mobile Banking or any otherelectronic/ non-cash mode of payment.3. Valid Identity proof is any of the following: Aadhaar Card, Driving License, VoterID Card, Pass Port, NREGA Card, PAN Card, Identity Card Issued by GovernmentDepartment, Public Sector Unit to its Sta↵.4. You may approach the control roomof RBI on Telephone Nos 022-22602201 226029445. The date for submission of annual life certificate has been extended to January 15, 2017from November for all government pensioners6. As of December 15, 2016, specified bank notes of only Rs. 500 can no longer be used forthe following: Government hospitals and pharmacies, railway and government bus tickets,consumer cooperative stores, government and court fees, government School fees, mobiletop-ups, milk booths, crematoria and burial grounds, LPG gas cylinders, ArchaelogicalSurvey of India monuments, utilities, toll payments

1


Appendix C. Example Pamphlet Excerpts

(a) Front

(b) Back

Figure C.1. Short pamphlet (2 facts)


(a) Front

(b) Page 1/8

(c) Page 2/8

Figure C.2. Long pamphlet (24 facts)


Appendix D. Proof of Proposition

First let us analyze CK:BC. Suppose that for infinitely many values of ⁄, therewere a di�erence bigger than ” > 0 between

PA(a = H | d = S) and PA(a = L | d = S).

Then as ⁄ ø Œ, the probabilty of a realization V S ≠ V NS exceeding the reputationalcost would tend to 0, a contradiction. This, along with P1, shows that as ⁄ ø Œ, theprobability ‡(1; CK:BC) tends to 0.

Now let us fix a large ⁄ and analyze the other inequalities. Under CK:Seed ourassumption that it is common knowledge that xD = 0 turns o� signaling concernsand so by P3 everybody seeks.

Thus it su�ces to argue that ‡(0; No-CK) < ‡(1; No-CK). The expected signalingcost is O(Á), where Á is defined to be D’s expectation of A’s subjective probabilitythat xD = 1. Under B1, Á = o(—) when xD = 1. However, the probability that thereis information to be gained when xD = 1 is �(—). It follows, given P3, that for anysmall —, the mass of types seeking is �(—). On the other hand, if xD = 0, then Dconsiders it very unlikely that information is out there (O(—2)) and so given P3 it canbe ensured that the mass of types seeking is O(—2) as well.


Appendix E. Supply Effects: Information as a Public Good

The core model of Chandrasekhar, Golub, and Yang (2017) and its applicationto our setting focuses on seeking e�ort or endogenous participation in learning. Adi�erent kind of explanation focuses on the e�ort of those informed to understand,filter, and communicate the information in a useful way to others. The simplestframework to capture this is a model of public goods provision and free-riding. Thisclass of model has been studied extensively in a development context, and we rely onarguments from Banerjee, Iyer, and Somanathan (2007) to explain why supply-sidee�ects are unlikely to explain our results.

A robust point within such public goods models is that enlarging the set of peoplewho are able to provide a public good should not, in equilibrium, reduce its aggregateprovision. Indeed, if anything provision should slightly increase, which is contrary toour empirical results.

For a simple model, consider a situation where those initially given informationhave the opportunity to provide the public good of processing and disseminating it toothers. There are n agents, and each of those informed believes that k in total are ableto contribute. Every i who has information invests an e�ort zi Ø 0 in transmitting.Their payo�s are given by

Ui(z1, . . . , zn) = V

Aÿ

i

zi

B

≠ czi.

Here V is an increasing, smooth function with V Õ(z) tending to 0 at large argumentsz, and c > 0 is a cost parameter. Those who are unable to contribute are con-strained to zi = 0 and are passive. The key fact, which is formalized for instance byBanerjee, Iyer, and Somanathan (2007), is that at any equilibrium with some peoplecontribution, for those contributing we have

(E.1) V ÕA

ÿ

i

zi

B

= c,

so the aggregate level of contribution cannot depend on n or k. The intuition is simple:the free-riding problem is self-limiting, at least in the sense of aggregate (though notper-person) provision. If more agents try to free-ride, then others have more reasonto provide the good. A similar force is present in the network model of Galeotti andGoyal (2010), in which everyone has access to the public good in equilibrium.

If agents have a private benefit term in their utility function, vi(zi), where v isincreasing and vÕ(zi) > c for zi œ [0, ”), then as long as there are su�ciently many


agents who can provide the public good, the amount provided will be at least k”—alower bound which is increasing in k. A similar argument applies if only some agentshave such a v term.

Thus, natural public goods theories do not predict a decrease in the amount of over-all provision, and thus in overall learning, as k (the number of potential providers)increases. One can, of course, elaborate these models with stochastic k and idiosyn-cratic ci, but the basic intuition described above is quite robust.

One further supply-side e�ect to consider is one of social obligation. If the seedsare publicly “deputized,” as they are in the CK treatment, each may face strongerincentives to provide information relative to a situation in which provision opportu-nities are di�use. Though this is outside a basic public goods model, our evidence onseed e�ort does not support this hypothesis.

Application to Experiment. The number of people, k, who can contribute is eitherk = 5 or k = n. Under common knowledge, this matches up with the beliefs agentshold, so in this sense the simple model is faithful to the experiment. Thus, the basicpublic goods theory predicts (contrary to the demand-side theory) that holding CKfixed and moving from Seed to Broadcast should not hurt aggregate provision.

When common knowledge is not present, agents will have beliefs about k. Butas long as their beliefs about k are reasonably consistent (e.g., agents have commonpriors about it), the essence of the above argument goes through: a stochastic versionof (E.1) still holds, and changes in beliefs about k alone should not lead to large swingsin provision.

This model is inconsistent with our empirical findings for several reasons. First,aggregate provision of e�ort cannot decline, as established above. If the number ofpeople a typical subject in our random sample conversed with measures conversationale�ort, this means that the number of conversations for the average person must notdecline. Column 1 of Table 5 shows that, conditional on common knowledge, goingfrom k = 5 to k = n corresponds to a 61% decline in the number of conversations (p =0.029), which means that aggregate contribution to conversations must be decreasing.

Second, the model suggests that the amount of value being generated cannot de-cline, since after all otherwise a given individual would have an incentive to putin some more e�ort to gain more marginal benefit. Here, we can measure this ei-ther through knowledge or choice quality. Turning to Table 6, recall that columns 1(for knowledge) and 2 (for choice) show robust declines in aggregate social learning


and quality of choice when we go from k = 5 to k = n under common knowledge(p = 0.0621 and p = 0.104).


Appendix F. Heterogeneity by length of information

We now look at the interaction of our core treatment cells with the amount ofinformation in the pamphlet. Whether this should accelerate or dampen the e�ect ofgoing to common knowledge in a given information delivery system depends on thedetails of the model and therefore becomes an empirical question.

To see why, consider the case of (Broadcast, CK) and now imagine comparing aworld in which only two facts are given as compared to a world where a lengthypamphlet of 24 facts is given. What matters is how the type-specific marginal valueof information distributions, FH and FL, move when we go from a short set of factsto a long set of facts. Assume for the moment that the cost of figuring out whichof the 24 facts are useful, or coordinating on the same topic of conversation outof the now 24 possibilities, is very high no matter if the individual is a high orlow ability type. In this case, the scope for signaling reduces, and therefore goingfrom (Broadcast, No CK, Long) to (Broadcast, CK, Long) should generate less of areduction in endogenous participation in social learning than going from (Broadcast,No CK, Short) to (Broadcast, CK, Short). Now on the other hand, if it was veryeasy for high ability types to figure out what is useful, but the task was arduous forlow ability types, then scope for signaling could actually increase.

Turning to seeding, observe that in seeding with or without common knowledge,the length of the information is not commonly known either way. So, long sets offacts should likely have no e�ect on endogenous participation.

We now turn to the data in Table E.1 to look at how going from two to 24 factsdi�erentially impacts the e�ects of interest. For the most part the e�ect is noisy, andthere is no di�erential e�ect. The one plausible finding is that going from (Broadcast,No CK) to (Broadcast, CK) is less of a deterrent to purposeful conversations (p =0.15) when the facts are long. If this is to be taken seriously, minding the caveat thatfor overall conversations this e�ect is not distinguishable from zero (p = 0.251), itis evidence in favor of the idea that sorting through the 24 facts or deciding whichtopic to coordinate on and converse about is costly enough for both ability typesthat the signaling motive is dampened by the longer list. Said di�erently, it is, ifanything, consistent with the story that it is much less likely for someone to go askabout information when it is known that they have received two facts, than when itis known that they received a lengthy booklet of facts.

Table E.2 repeats the same exercise now turning to knowledge and choice. Of noteis that a similar pattern is true here. There is mostly no detectable e�ect. But if we


Table E.1. Conversations: Length interactions

(1) (2) (3)OLS OLS OLS

Volume of # incidental # purposefulVARIABLES conversations conversations conversations

CK 0.825 0.693 0.132(0.496) (0.412) (0.163)[0.0982] [0.0937] [0.421]

Broadcast 0.963 0.665 0.297(0.545) (0.481) (0.219)[0.0787] [0.168] [0.175]

Long -0.0939 -0.00127 -0.0926(0.372) (0.330) (0.130)[0.801] [0.997] [0.478]

Broadcast ◊ CK -2.212 -1.614 -0.599(0.735) (0.626) (0.264)

[0.00296] [0.0107] [0.0244]CK ◊ Long -0.372 -0.485 0.113

(0.562) (0.480) (0.194)[0.508] [0.313] [0.560]

Broadcast ◊ Long -0.563 -0.319 -0.244(0.680) (0.616) (0.233)[0.408] [0.605] [0.295]

Broadcast ◊ CK ◊ Long 1.448 1.006 0.442(0.809) (0.733) (0.281)[0.0752] [0.172] [0.118]

Observations 1,078 1,078 1,078Seed, No CK, Short Mean 0.523 0.385 0.138CK + BC ◊ CK = 0 p-val 0.00573 0.0275 0.0365BC + BC ◊ CK = 0 p-val 0.0170 0.0259 0.0602Long + CK ◊ Long + BC ◊ Long + BC ◊ CK ◊ Long = 0 0.251 0.520 0.155Notes: All columns control for randomization strata (subdistrict) fixed e�ects. They also control for dateand time of entry into the village, caste category of the treatment hamlet and distance from the village toan urban center. Respondent-level controls include age, gender, literacy and potential seed status. Standarderrors (clustered at the village-level) are reported in parentheses and p-values are reported in brackets.

had to guess, at p = 0.5 for both outcomes, it suggests that perhaps introducing CKto the broadcast cell has less of a detrimental e�ect on both knowledge and choicequality. This is extremely noisy, speculative evidence that suggests if anything, astigma-like e�ect operates more when there are only two facts.


Table E.2. Knowledge and choice: Length interactions

(1) (2)OLS OLS

VARIABLES Knowledge Error Chose 500

CK -0.0215 0.0542(0.0162) (0.0404)[0.185] [0.181]

Broadcast -0.0264 0.0804(0.0169) (0.0361)[0.121] [0.0269]

Long 0.0131 -0.00591(0.0174) (0.0300)[0.451] [0.844]

Broadcast ◊ CK 0.0537 -0.144(0.0247) (0.0556)[0.0312] [0.0104]

CK ◊ Long -0.0167 -0.0144(0.0255) (0.0508)[0.513] [0.777]

Broadcast ◊ Long 0.000655 -0.0284(0.0262) (0.0548)[0.980] [0.605]

Broadcast ◊ CK ◊ Long -0.00862 0.0696(0.0383) (0.0785)[0.822] [0.376]

Observations 1,082 1,067Seed, No CK, Short Mean 0.436 0.0374CK + BC ◊ CK = 0 p-val 0.0919 0.0141BC + BC ◊ CK = 0 p-val 0.120 0.133Long + CK ◊ Long + BC ◊ Long + BC ◊ CK ◊ Long = 0 0.532 0.550Notes: All columns control for randomization strata (subdistrict) fixed e�ects. They alsocontrol for date and time of entry into the village, caste category of the treatment hamletand distance from the village to an urban center. Respondent-level controls include age, gen-der, literacy and potential seed status. Standard errors (clustered at the village-level) arereported in parentheses and p-values are reported in brackets.

information delivery under endogenous communication anushka chawla, devika lakhote, alokik mishra,...

Documents